Research Summary: Decision-Centered Evaluation of Machine Learning Poverty Maps Using Mobile Phone and Satellite Data

New research evaluates machine learning (ML) models for poverty mapping in Sri Lanka, utilizing mobile phone and satellite data as alternatives to costly and infrequent traditional surveys. The study assesses the effectiveness of these models in identifying the poorest administrative units, particularly under budget constraints, highlighting the potential for improved resource allocation in poverty alleviation efforts.

Why it matters

This research is strategically important because it demonstrates a technology-driven approach to a critical social and economic challenge: identifying poverty. Accurate and timely poverty mapping can significantly improve the efficiency and effectiveness of resource allocation for social welfare programs, enabling more targeted interventions and potentially reducing operational costs associated with traditional data collection.

Key insights

• Traditional household surveys and censuses for poverty identification are costly and infrequent.
• Machine learning, using mobile phone call detail records (CDRs) and satellite remote sensing (RS) data, offers an alternative for poverty estimation.
• The research evaluates ML poverty maps for 13,985 administrative divisions in Sri Lanka.
• Evaluation criteria include the recovery rate of the poorest administrative units and comparison of validation methods (random vs. spatially grouped).
• The study also examines prediction errors in socioeconomically atypical communities.
• A Random Forest model, using a combination of CDRs, RS, and CNN-derived Landsat 8 embeddings, was applied against a census-derived asset index.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.23805